Sixteen recently published pediatric glioblastoma RNA-seq samples were chosen as the first training data (International Cancer Genome Consortium PedBrain Tumor Project, 2016). Fusion gene candidates in these 16 samples were first identified by tools SOAPfuse and TopHat2-Fusion (Kim et al., 2013). High-confidence candidates were then filtered for common artifacts and by visual inspection of fusion break points of exons between two fusion partners (International Cancer Genome Consortium PedBrain Tumor Project, 2016). In total, 40 fusion genes were successfully verified by RT-PCR among the 16 samples. The first training data was mainly used to identify and select features from deFuse reports.

The second training data contains 96 RNA-seq samples from seven studies, including pilocytic astrocytoma (n = 7; Jones et al., 2013), thyroid cancer (n = 5; Ricarte-Filho et al., 2013), glioblastoma (n = 47; Bao et al., 2014), lung adenocarcinoma (n = 28; Seo et al., 2012), ependymoma (n = 7; Pajtler et al., 2015), lung cancer liver metastasis (n = 1; Ju et al., 2012), and biphenotypic sinonasal sarcoma (n = 1; Wang et al., 2014). Those samples, including 126 experimentally verified fusion genes, were used to optimize the score weights of varied features.

A published study of early-onset prostate cancer including 11 RNA-seq samples were chosen for validation in silico (Weischenfeldt et al., 2013) and three primary breast cancer samples were used for experimental validation.

Despite a prominent role for oncogenic gene fusions in multiple cancer types, it is relatively rare for the exact same fusion to be detected across multiple, unrelated tumor entities. Fusions identified in multiple samples from different tumor entities based on currently available fusion detection tools are therefore mostly considered to be of high false positive rate. This high false positive prediction may be due to genomic complexity such as repeat regions or mapping artifacts in the alignment step. In total, 171 paired-end RNA-seq samples from 15 different entities were used to generate an artifact list of fusions identified in multiple samples from several different entities (Figure 1). To increase the sensitivity, a small number of verified fusions were manually excluded from the artifact list. We aim to assign high-confidence fusion genes a score between 8 and 10, and consider fusions contained in the artifact list to be of high false positive rate. confFuse therefore assigns a negative score (−6) to those fusions in the artifact list in order to rank them outside of the range of confident predictions.

When taking fusion candidates identified by deFuse (version v0.6.1) in no less than three entities (recurrence ≥ 3), 2190 fusions were included in the artifact list (Figure 2). For candidates identified in ≥4 and ≥5 entities, there are 1,409 and 995 fusions in the artifact list, respectively. In this study, we chose a threshold of three entities for the final artifact list. Of note, a small number of additional putative artifacts are still identified with each increase in the number of different tumor types, suggesting that accuracy could be further improved by increasing the complexity of the data set used to generate the artifact list.

In total, 64% (5881/9169) of putative fusion transcripts (n = 9169) identified in the second training data were found in the artifact list (n = 2190). Among them, 62.3% (3666/5881) were fusions from adjacent genes and 91.5% (5378/5881) were identified by deFuse as likely being a product of alternative splicing.

One of the most important features supporting a true fusion event is the number of split reads and spanning reads. Since this is related not just to mapping performance, but also to fusion gene expression levels and sequencing depth, we found that setting a simple threshold on the number of split and spanning reads could not best distinguish true and false positive predictions. For example, a true fusion gene with low expression and low coverage sequencing depth may have only a few detectable split and spanning reads. A false positive fusion gene may have a large number of reads due to mapping artifacts and/or unreliable reads aligned to multiple genomic locations. Comparing verified fusions with all initial calls, the distribution of number of split reads and spanning reads between them is similar, of which the majority are of low read numbers (Figure 3). Most of the verified fusions in the first training data have <200 split reads and 50 spanning reads. A threshold purely on the number of split and spanning reads therefore cannot distinguish true and false positive fusion predictions.

In addition, the mapping quality of reads should also be considered. Some spanning reads can be aligned to more than one genomic position, indicating reads of low mapping quality which do not reliably support a fusion event. Breakpoint homology is the number of nucleotides near the fusion break point which can map equally well to both fusion partners, with very high breakpoint homology therefore suggesting more ambiguous support for a fusion event. Most of the verified fusions contain <10 homologous bases at the fusion breakpoint (Figure 3). confFuse therefore assigns a negative score (−1) when breakpoint homology is ≥10. If spanning reads are mapped on a repeat region, it is difficult to identify where they are originally from, and thus those spanning reads are considered as low mapping quality. Therefore, confFuse assigns a negative score to those fusions with the majority of spanning reads aligned on repeat regions, e.g., −0.5 score for fusions of 80% up to 90% of spanning reads aligned on a repeat region and −1 score for those fusions of >90% (Figure 3; Table S1).

Fusion genes with different fusion transcripts (i.e., splice variants) in the same sample may be of high true positive rate, especially those fusion transcripts with a high count of split reads and spanning reads. We observed that deFuse sometimes reports multiple fusion transcripts for the same pair of fusion partner genes (occurrence of the same fusion pairs) indicating high probability of true fusion events. confFuse thus considers the multiple fusion transcripts from a fusion gene as a positive factor supporting a true fusion event.

By combining with the occurrence of the same fusion pairs, the number of supporting reads, mapping quality of reads, possible mapping artifacts and other fusion structure related features mentioned below, confFuse also assigns a positive score from 0.5 to 2.5 to fusions with a high number of split and uniquely mapped spanning reads or a negative score from −0.5 to −2.0 otherwise, such as −1.5 score when all the spanning reads are mapped on more than one genomic location (Table S1).

Two adjacent genes in the same orientation may give rise to an apparent fusion due to read-through transcription or aberrant splicing rather than genomic rearrangement. Although some may acquire novel function, the vast majority are expected to be false positives in terms of their biological relevance. deFuse reports an altsplice feature, indicating that a fusion may arise from alternative splicing between adjacent genes. In the first training data, verified fusion genes do not contain any read-through or alternative splicing events (Figure 4). More than 75% of initially identified fusions are, however, from an alternative splicing event. Therefore, confFuse takes those fusions with read-through or alternative splicing as high false positive fusion candidates by assigning a negative score (−4) in order to separate them from high confidence candidates. An adjacent gene fusion ESR1:CCDC170 was reported from 22 of 990 tumor samples (Veeraraghavan et al., 2014), showing the possibility of true fusions from adjacent genes. Fusion candidates from adjacent genes but without a read-through or altsplice tag are therefore given a reduced penalty of only −0.5. Furthermore, a higher ratio of inter- rather than intra-chromosomal fusions were detected in the verified fusions comparing with all identified fusions (Figure 4), and confFuse therefore also assigns a modest negative score (−0.5) for intrachromosomal predictions (users can reset the score weight to zero if distance between fusion gene partners is large).

True oncogenic fusions typically preserve the open reading frame in order to form a functional fusion protein, and the precise location of a fusion breakpoint point plays a critical role in demonstrating evidence supporting true positive fusions. When the location of a fusion splicing point is at a known exon boundary, such a fusion is more likely to be a true positive. Comparing with all identified fusions, we observed higher ratios of verified fusions preserving an open reading frame and showing fusion splicing points at exon boundaries (Figure 4). confFuse assigns a negative score to fusions with non-detected open reading frame (−1 score) and with splicing point not at an exon boundary (−1.5 score). It is also more likely to be of low biological interest when a break point is located downstream of the 3′ fusion partner because such predictive fusions may not have biological function or may arise from mapping artifacts. confFuse takes those fusions as low-confidence ones by assigning a negative score (−4).

In the first training data (n = 16), 77.5% (31/40) of verified fusions were scored ≥8 by confFuse, 15% (6/40) is of 6 ≤ score <8, and 2.5% (1/40) is <6 (Table S2). In total, 8,083 fusion gene candidates (9,169 putative transcripts) from the second training data (n = 96) were identified by deFuse using default settings, of which 126 fusions were previously validated by RT-PCR (Table S3). confFuse called 301 high-confidence fusion genes (score ≥ 8, 301/8,083, 3.7%). Among the 301 fusions were 108 of the 126 validated fusions, resulting in a recovery rate of 85.7% (108/126). The remaining previously validated fusions were either scored <8 (n = 5) or were not detected or were already filtered by default deFuse parameters prior to application of confFuse (n = 13; Table S4). The correlation between recovery rate and confFuse score in the second training data is given in Figure 5. As annotation features such as read-through are not provided by fusionMap and soapFuse, the number of supporting reads was chosen to compare recovery rate. As the threshold for the required minimum number of supporting reads increases, the recovery rate by soapFuse and fusionMap decreases (Figure S1, Tables S9, S10). For fusions with more than five supporting reads, fusionMap predicted 964 fusions, of which 101 were previously verified, resulting in a recovery rate ~80% (101/126); soapFuse predicted 562 fusions, of which 65 were verified before, resulting in a recovery rate ~51.6% (65/126). Thus, setting a threshold only on the number of supporting reads cannot retain a high recovery rate and results in decreasing sensitivity.

The distribution of deFuse's own probability score was compared with our confFuse confidence score for all putative fusion transcripts in the second training data (Figure 6). Notably, there are many putative transcripts with a high deFuse probability which were assigned a low score (~ −8) by confFuse, of which most are in the artifact list or of alternative splicing feature. None of these are in the list of 126 known validated fusions in the second training data. Most of the verified fusions (108/126) are located in the range of confFuse confidence score no less than eight. It demonstrates that confFuse is able to identify the verified fusions among hundreds of putative fusions with high deFuse probability.

Comparing across tools, fusionMap, deFuse (default probability score threshold = 0.5), deFuse-0.81 (deFuse with probability score threshold set to ≥0.81, as used in McPherson et al., 2011), soapFuse, confFuse-6.5 (score ≥ 6.5) and confFuse-8 (high-confidence candidates scored ≥ 8) showed recovery rates of 91.3, 89.7, 84.9, 73, 88.1, and 85.7% respectively for the 126 validated fusions (Figure 7; Tables S4–S6, S9, S10), indicating confFuse can dramatically reduce the number of candidates (from 8,083 to only 301) without compromising detection accuracy when compared with other available tools.

To evaluate the accuracy of high-confidence candidate predictions (score ≥ 8), three primary breast tumor samples were sequenced to generate paired-end RNA-seq data. In total, deFuse predicted 1,026 fusion genes in the three samples, of which 18 scored ≥8 by confFuse. All 18 high-confidence candidates were validated with RT-PCR followed by Sanger sequencing, resulting in 100% validation rate (Figure 8; Table S7). To the best of our knowledge, the 18 novel fusion genes haven't been validated before. Interestingly, one of them (QKI:PACRG) was predicted in three of 1,019 breast cancer samples from TCGA (www.tumorfusions.org), indicating that QKI:PACRG may be a novel recurrent fusion in breast cancer. QKI can suppress cell proliferation and prevent inappropriate activation of the Notch signaling pathway in lung cancer (Zong et al., 2014) and PACRG is an evolutionarily conserved protein with currently unclear function (Dawe et al., 2005). Furthermore, we randomly chose some candidates scored <8 for validation. Nine of 13 candidates (6.5 ≤ score ≤ 7.5, medium confidence) and three of 10 candidates (0.5 ≤ score ≤ 6, low confidence) were experimentally verified (Table S7). Most of the genes involved in verified fusions were in the expression level of FPKM <50 (median≈6.44; Figure S2), indicating that it is not only very highly-expressed candidates being detected but also lowly-expressed ones.

To find out whether or not those verified fusions are individual tumor specific rather than simply artifacts of pan-breast tumor expression, the primers of verified fusions in each sample were used for validation in the other two breast primary tumor samples as control. In total, primers (Untergasser et al., 2012) of 14 verified fusions were used, of which 13 showed true negative in two control samples and one showed a false negative in one control sample DR8V (Figure S3). This “false negative” fusion (CD3D:TOM1L2) in sample DR8V shows a much weaker band in gel compared with the one in sample AC72 where the fusion was true positive, and appears as a double band of different size to AC72, possibly indicating an unspecific PCR product.

In addition, 11 published early-onset prostate cancer samples were used as an in silico validation dataset. Eight hundred and forty-nine fusion genes were called by deFuse, 24 of which were identified as high-confidence fusion (score ≥ 8) by confFuse. The well-known E26 transformation-specific (ETS) fusions were detected by confFuse in 10 of 11 samples, which are the same as previously published results (Weischenfeldt et al., 2013). Among the 24 fusions, 17 were confirmed by DNA-seq, FISH validation or known ETS fusions, resulting in a ~70% (17/24) recovery rate (Table S8).

Overall, confFuse can classify fusion candidates into three groups, namely high-confidence (8 ≤ score ≤ 10), medium-confidence (6.5 ≤ score ≤ 7.5) and low-confidence (score ≤ 6) fusions, which makes the users more convenient to prioritize candidates for validations. Not only novel and biologically relevant fusions can be identified by confFuse, but also well-known fusions across tumor entities can be detected by confFuse.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.